GASC-Net: A Geospatial information-assisted network for ship classification

Abstract

Recently ship classification in optical images has received increasing interest, which can be categorized as coarse-grained classification, fine-grained classification, and instance-level classification according to the scope of the sort. Due to the influence of cloud occlusion, insufficient lighting, etc., it is challenging for finer classification when only images are used. In this paper, geospatial information is introduced into ship classification for different level classifications. A geospatial information-assisted ship classification network named GASC-Net is proposed. GASC-Net consists of a feature extractor backbone, a Siamese Position Encoding (SPE) module, and a Geographical Position Fusion Attention (GPFA) module. The longitude and latitude position information of ships is sent to SPE module for position encoding. The position-coding information is combined with image features via GPFA, which GPFA fuses positional encoding information into image features by channel attention. Extensive experiments are taken on a Geospatial Ship dataset, showing that GASC-Net can obtain state-of-the-art performance.